Source-linked AI summary
Your click decides your fate: Inferring Information Processing and Attrition Behavior from MOOC Video Clickstream Interactions
Tanmay Sinha, Patrick Jermann, Nan Li, Pierre Dillenbourg
TL;DR
MOOC research needs better ways to understand participation, engagement, and attrition in large-scale video-based courses. This paper converts clickstreams into cognitively motivated behaviors and an Information Processing Index, reporting associations between higher processing, rewatching, video watching, and lower dropout likelihood.
Problem
MOOCs face unsatisfactory outcomes and attrition, while researchers need to understand how learners allocate attention and develop participation patterns during course-material interaction.
Method
The paper operationalizes video clickstreams as higher-level behavioral categories and constructs a cognitive-psychology-based Information Processing Index for learner-partition and trajectory analysis.
Results
37% lower dropout likelihood accompanies a one-standard-deviation-greater IPI, while high rewatching and greater video coverage correspond to 33% and 37% lower dropout likelihood, respectively.
Takeaways & Limitations
The metric can help instructors reason about learner behavior, identify video segments needing editing, and potentially provide real-time or adaptive support.
Takeaways & Limitations
Because the platform does not log mid-video departures, the study includes only sessions containing video-end pauses, leaving true engagement time unknown.
Abstract
from arXiv · showhide
In this work, we explore video lecture interaction in Massive Open Online Courses (MOOCs), which is central to student learning experience on these educational platforms. As a research contribution, we operationalize video lecture clickstreams of students into cognitively plausible higher level behaviors, and construct a quantitative information processing index, which can aid instructors to better understand MOOC hurdles and reason about unsatisfactory learning outcomes. Our results illustrate how such a metric inspired by cognitive psychology can help answer critical questions regarding students' engagement, their future click interactions and participation trajectories that lead to in-video & course dropouts. Implications for research and practice are discussed
1 Introduction
MOOCs broaden access to self-paced learning, but standardized instruction, limited individualized attention, and fluctuating motivation contribute to attrition and unsatisfactory outcomes. The paper addresses this challenge by analyzing video clickstreams as signals of engagement, information processing, and dropout trajectories.
- Motivation: MOOCs connect large numbers of learners and let them engage with instruction at their own pace, yet learning outcomes remain unsatisfactory.The introduction frames MOOCs as scalable and potentially transformative while noting skepticism about their productiveness.
- Motivation: Standardized delivery, large classes, limited face-to-face interaction, and high student-teacher ratios make individualized guidance difficult and contribute to fluctuating motivation and attrition.The paper identifies personalization and regulation-of-learning support as unmet needs.
- Research gap: Understanding how participation ecologies and attention allocation develop during course-material interaction is presented as a prerequisite for improving MOOC support.The authors distinguish this need from prior emphasis on discussion forums and assignment grading.
- Video lectures: Video lectures are a primary MOOC activity and serve as gateways for course concepts, assignments, and discussion participation.At EPFL, 75% of registrants attended, with 50% primarily watching video lectures.
- Contribution: The study develops hierarchical clickstream methods and an information processing index to predict engagement, future clicks, and in-video or course dropout.The analysis is rooted in cognitive psychology and aims to make low-level behaviors useful for predicting higher-level outcomes.
2 Study Context
The study analyzes interaction logs from a 2012 Coursera introductory programming MOOC, covering 48 video lectures and click-level playback behavior. Because the platform does not record mid-video departure reliably, the analysis includes only sessions ending with video-end pauses.
- Dataset: The dataset comes from Coursera’s 2012 “Functional Programming in Scala” introductory programming MOOC, comprising 48 video lectures and 10 Gb of JSON data.The logs register play, pause, seeking, scrolling, rate changes, playback rate, and time spent playing.
- Data limitation: The platform does not log whether students leave videos midstream, leaving true engagement time unknown.This is an explicit measurement limitation of the study context.
- Data limitation: The analysis retains only video sessions containing video-end pauses, yielding 222021 sessions from 21952 students.The filtering rule is used to avoid biased engagement-time data.
3 Operationalizing the Clickstream
The paper transforms raw video-player clicks into interpretable behavioral categories, weights those categories using fuzzy matching, and aggregates them into session-level behavioral vectors. It then maps these behaviors onto a cognitive information-processing hierarchy to construct the Information Processing Index (IPI), while recognizing that processing depends on both learner and video context.
- Level 1: Operations: Raw player events are encoded into eight categories, including play, pause, seeking, scrolling, and fast or slow rate changes.Pairs of seeks within less than 1 second are grouped as scrolling, and rate changes are classified by comparing prior and new playback rates.
- Level 1: Operations: Level 1 concatenates encoded events into per-student, per-lecture video-watching state sequences.Examples include sequences such as PlPaSfSfPaSbPa.
- Level 2: Behavioral actions: Grouping fine-grained clicks into higher-level categories improves interpretability and noise tolerance compared with raw click representations.The authors argue that many fine-resolution sequences are semantically equivalent and difficult to interpret.
- Level 2: Behavioral actions: Maximum-frequency n-grams, with n=4 selected to limit overspecificity, are used to identify recurring click combinations.These n-grams support behavioral categories such as rewatching, skipping, fast watching, slow watching, and clear concept processing.
- Level 2: Behavioral actions: Fuzzy string matching assigns behavioral-pattern weights by comparing category patterns with each student’s full clickstream sequence.The approach is intended to tolerate several forms of noise in raw clickstream logs.
- Level 2: Behavioral actions: Category-level similarity weights are summed and discretized into High or Low values, producing a summarized behavioral vector for each viewing session.Equal-frequency discretization separates relatively positive and negative similarity values despite differing distance-metric ranges.
- Level 3: Information processing: The cognitive model treats video watching as joint control by the student’s viewing choices and the medium’s information-processing demands.Students allocate limited cognitive resources while expertise, perceived difficulty, and motivation affect processing across pipeline stages.
- Level 3: Information processing: The Information Processing Index assigns linear weights to behavioral actions according to a hierarchy from higher to lower information processing.The index operationalizes clickstream behavior through weighted behavioral categories, though the precise hierarchy rationale is deferred to prior work.
4 Validation Experiments
The paper validates a hierarchical clickstream analysis through experiments predicting engagement, next-click states, in-video dropout, and complete course dropout. It also tests whether the Information Processing Index (IPI) varies systematically across student participation groups and whether summarized behavioral features retain predictive utility.
- Validation Experiments: Machine-learning experiments predict engagement duration, next click states, in-video dropout, and complete course dropout from increasingly abstracted clickstream and trajectory features.The analyses use engagement, click sequences, behavioral action vectors, and trajectories of engagement, video play proportion, and IPI.
- Will you drop out of the video?: In-video dropout prediction uses clickstream n-grams, interaction proportions, engagement, the last click action, post-click time, and summarized behavioral action vectors.The dependent variable is binary in-video dropout, modeled with cost-sensitive L2-regularized logistic regression and 10-fold student-annotated cross-validation.
- Results: Summarized behavioral action vectors achieve nearly similar accuracy and kappa to raw clicks, while lowering false negative rates in Cases 1.B and 3.B.The result supports reasoning about meaningful video-viewing behaviors without examining continually occurring raw clicks.
- Results: 37% less likely dropout is associated with an IPI one standard deviation above average (Hazard ratio: 0.6367, p<0.001).The paper also reports lower dropout likelihood for higher rewatching and greater video-play proportion.
- Results: Average IPI is significantly higher for high-engagement students and non-in-video dropouts in video 4-6, with |z|=8.296 and |z|=22.54 respectively, both p<0.01.Across the full course, IPI distinguishes non-dropouts from dropouts (|z|=9.06, p<0.01) and active students from viewers (|z|=10.45, p<0.01).
5 Conclusion
The study lays a foundation for analyzing information processing during MOOC video interactions by applying a cognitive video-watching model. It develops an Information Processing Index (IPI) as a predictive operationalization for critical learner behaviors.
- The authors apply a cognitive video-watching model to explain the dynamic cognitive process underlying MOOC video clickstream interactions.
- The resulting IPI uses linear weight assignments to operationalize information processing and support predictions about critical learner behavior.
- IPI varies across student partitions because larger groups contain smaller substructures with similar click behaviors.
- Markov-based clustering could identify distinct interaction patterns within smaller student clusters and inform customized learning solutions.
- Student demographics could help distinguish latent influences on playback-speed choices, such as native-language differences versus engagement.